Klotski - Zhe Wu

Transcription

Klotski - Zhe Wu
Klotski: Reprioritizing Web Content to
Improve User Experience on Mobile Devices
Michael Butkiewicz♔, Daimeng Wang♔,
Zhe Wu♖, Harsha V. Madhyastha♖, Vyas Sekar♘
♔
UC Riverside
♖
University of Michigan
♘
CMU
Motivation: Slow Mobile Web
Motivation: Slow Mobile Web
2.5
sec
0.5
7.5
sec
12
sec
Motivation: Slow Mobile Web
2.5
sec
7.5
sec
12
sec
0.5
Slow page loads → Less users, Lost business
Wide Range of Existing Solutions
Wide Range of Existing Solutions
Compression
Caching
Wide Range of Existing Solutions
Mobile format
web pages
Compression
SPDY
Caching
Wide Range of Existing Solutions
Mobile format
web pages
Compression
SPDY
Faster
Networks
Better web
browsers
Caching
Wide Range of Existing Solutions
Mobile format
web pages
Compression
SPDY
Common Focus: Reduce Page LoadCaching
Times
Faster
Networks
Better web
browsers
Reducing Load Time is Not Enough
Rising Website Complexity
Reducing Load Time is Not Enough
Rising Website Complexity
Falling User Tolerance
Our Approach:
Dynamic Reprioritization
● Forseeable Future: High load times norm, not exception
Our Approach:
Dynamic Reprioritization
● Forseeable Future: High load times norm, not exception
● Reformulate the problem:
Our Approach:
Dynamic Reprioritization
● Forseeable Future: High load times norm, not exception
● Reformulate the problem:
How to reduce page load time?
Our Approach:
Dynamic Reprioritization
● Forseeable Future: High load times norm, not exception
● Reformulate the problem:
How to reduce page load time?
How to Prioritize the content most important to user?
Our Approach:
Dynamic Reprioritization
● Forseeable Future: High load times norm, not exception
● Reformulate the problem:
How to reduce page load time?
How to Prioritize the content most important to user?
○ Typical tolerance limit of 2-4 seconds
○ Deliver “high utility” resources within time budget
Our Approach:
Dynamic Reprioritization
● Forseeable Future: High load times norm, not exception
● Reformulate the problem:
How to reduce page load time?
How to Prioritize the content most important to user?
○ Typical tolerance limit of 2-4 seconds
○ Deliver “high utility” resources within time budget
● Our solution: Klotski proxy
○ No modifications to clients and web servers!
Klotski in Action!
Original Page Load at 3s
Klotski Page Load at 3s
Klotski: Idealized View
Web Server
Klotski
Proxy
Webpage Resource
Klotski: Idealized View
Web Server
Klotski
Proxy
Webpage Resource
Klotski: Idealized View
Web Server
Klotski
Proxy
High Priority Resource
Low Priority Resource
Klotski: Idealized View
Web Server
time budget
Klotski
Proxy
High Priority Resource
Low Priority Resource
Challenges with Idealized View
C1: Dynamic content + dependencies
Web Server
time budget
Klotski
Proxy
High Priority Resource
Low Priority Resource
Challenges with Idealized View
C1: Dynamic content + dependencies
C2: Fast selection of subset to prioritize
Web Server
time budget
Klotski
Proxy
High Priority Resource
Low Priority Resource
Challenges with Idealized View
C1: Dynamic content + dependencies
C2: Fast selection of subset to prioritize
C3: How long will a subset take?
Web Server
time budget
Klotski
Proxy
High Priority Resource
Low Priority Resource
Challenge 1:
Dynamic Content and Dependencies
Robust Dependency Knowledge
Image
A
requires
Javascript
A
Web Server
time budget
Klotski
Proxy
A
High Priority Resource
Low Priority Resource
Intuition: Page structure is stable
Dependency DAG
HTML
CSS
SCRIPT
IMAGE 2
IMAGE 1
IMAGE 3
● Prior work on static dependencies
○ E.g., WebProphet, WProf
Intuition: Page structure is stable
Dependency DAG
HTML
CSS
SCRIPT
IMAGE 2
IMAGE 1
IMAGE 3
● Prior work on static dependencies
○ E.g., WebProphet, WProf
● Problem: Dependencies not
reusable due to dynamic content
Intuition: Page structure is stable
Dependency DAG
HTML
CSS
SCRIPT
IMAGE 2
IMAGE 1
● Prior work on static dependencies
○ E.g., WebProphet, WProf
● Problem: Dependencies not
reusable due to dynamic content
IMAGE 3
● Our observation:
○ Nodes in DAG change
○ DAG structure largely stable
Intuition: Page structure is stable
Dependency DAG
HTML
CSS
SCRIPT
IMAGE 2
IMAGE 1
● Prior work on static dependencies
○ E.g., WebProphet, WProf
● Problem: Dependencies not
reusable due to dynamic content
IMAGE 3
● Our observation:
○ Nodes in DAG change
Load page repeatedly to generate fingerprint:
○ DAG structure largely stable
● DAG structure with a URL pattern at every node
● Pattern generalizes URL of dynamic resources
Learn URL Patterns
● Generalize known prior URLs of a dynamic resource
foo.com/SG39HZ78/a.js
foo.com/SHFS2732/a.js
→ foo.com/*/a.js
Learn URL Patterns
● Generalize known prior URLs of a dynamic resource
foo.com/SG39HZ78/a.js
foo.com/SHFS2732/a.js
→ foo.com/*/a.js
● 3 Cases = 90% of Replacements:
○ Single token in URL changes
○ Only URL argument changes: www.site.org/a.js?FOO=1...
○ CDN node name: {CDN2.bar.com/x.jpg, CDN5.bar.com/x.jpg}
Identify Resource Replacements
● Capture set of prior URLs:
Track Replacement of Resource
over multiple page loads
Identify Resource Replacements
R1
R2
● Capture set of prior URLs:
Track Replacement of Resource
over multiple page loads
Identify Resource Replacements
R1
R2
● Capture set of prior URLs:
Track Replacement of Resource
over multiple page loads
● Combination of techniques:
○ Match position in DAG
dependency structure
○ Identical position on screen
○ Identical reference in source
Challenge 2:
Optimal Prioritization Schedule
Which subset
to prioritize
and how?
Web Server
time budget
Klotski
Proxy
High Priority Resource
Low Priority Resource
Select Subset of High Utility Resources
Dependency
structure in
fingerprint
Select Subset of High Utility Resources
Dependency
structure in
fingerprint
Select Subset of High Utility Resources
Dependency
structure in
fingerprint
DAG Cut
● Reduces to knapsack with dependencies: NP-Hard
● Apply greedy heuristic
Prioritizing High-Utility Resources
● Static URLs: Use SPDY PUSH to pre-emptively deliver
as soon as main HTML is requested
● Dynamic URLs: Prioritize delivery if match with regular
expression of a selected resource
Challenge 3:
Estimating Load Times in the “Wild”
Will subset load
within budget?
X
Web Server
time budget
Klotski
Proxy
High Priority Resource
Low Priority Resource
Seemingly Natural Non-Solutions
Model: bytes, #requests, ..?
Use Apriori Load Times?
Web Server
time budget
Klotski
Proxy
Model: Too simplistic to capture browser effects
High Priority Resource
Apriori Loads: No longer valid with reprioritization
Low Priority Resource
Cannot capture diversity in client conditions
Intuition Behind Klotski Estimator
Web Server
time budget
Klotski
Proxy
Bottleneck = Client-Proxy Link (e.g, 4G)
Intuition Behind Klotski Estimator
Web Server
time budget
Klotski
Proxy
Build Fluid Model Simulation
● Proxy as work conserving scheduler
● Priorities/Dependencies, fairly shared bw for concurrent transfers
● Some subtle issues: PUSH, client processing delays
Klotski: System Architecture
Web Servers
Klotski Service
BackEnd
Resource
Selection
FrontEnd
Fingerprint
Generator
Load Time
Estimator
Clients
Klotski: Experimental Results
Data Set
Websites
● 50 Random From Alexa Top 200
Mobile Device
● Android Smartphone
● 4G Connection
● Google Chrome
Paper Only Results
●
●
●
●
Desktop PC + Ethernet
Smartphone + Full Website
User study on utility preferences
Resource churn over time
Klotski Improves User Experience
Klotski Improves User Experience
Within the first
X seconds
Klotski Improves User Experience
% of site's high util
resources loaded
Within the first
X seconds
Klotski Improves User Experience
50 Websites
Klotski Improves User Experience
50 Websites
75%
50%
25%
Klotski Improves User Experience
0.25
Median site
25% ATF
in 2 secs
Klotski Improves User Experience
Klotski Improves User Experience
35%
utility
gain
Klotski Improves User Experience
2x Util
25% ➙ 60%
High Utility Gain For Diverse Users
High Utility Gain For Diverse Users
User
Preferences
High Utility Gain For Diverse Users
Utility Gained
at 2 Seconds
User
Preferences
High Utility Gain For Diverse Users
Conclusions
● Mobile web continues to be a pain point
○ Focus on load time alone is likely insufficient
● Instead we focus on dynamic reprioritization
● Key challenges we address:
○ dynamic dependency representation
○ fast resource selection
○ load time estimation
● Klotski greatly improves user experience
○ for diverse preferences